Ensure the graphviz filenames are reproducible

[lamby]

Whilst working on the Reproducible Builds effort, we noticed that sphinx could generate output that is not reproducible.

In particular, the graphviz extension module would construct filenames based on, inter alia, the contents of the options dictionary.

As this contained the absolute build path of the source file embedded in the docname variable this meant that builds of documentation were not independent of where on a filesystem they were built from

Example filenames might be:

  -  html/_images/graphviz-9e71e0f9ba91d0842b51211b676ec4adb7e7afb8.png
  +  html/_images/graphviz-6241bbfd7ac6bd4e2ad9af451ab0dfb8719988d2.png

We fix this by limiting how much of the docname variable ends up in the final constructed filename; I assume there is a good reason for including the options dictionary in the first place, otherwise we could simply omit it.

This was originally filed in @Debian as #921513.

[codecov]

Codecov Report

Merging #6028 into master will increase coverage by 0.02%.
The diff coverage is 0%.

@@            Coverage Diff             @@
##           master    #6028      +/-   ##
==========================================
+ Coverage   82.91%   82.93%   +0.02%     
==========================================
  Files         271      269       -2     
  Lines       39086    39054      -32     
  Branches     5847     5844       -3     
==========================================
- Hits        32408    32391      -17     
+ Misses       5330     5318      -12     
+ Partials     1348     1345       -3

Impacted Files	Coverage Δ
sphinx/ext/graphviz.py	32.41% <0%> (-0.26%)	⬇️
sphinx/__init__.py
sphinx/make_mode.py

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update dc0120b...1a4362b. Read the comment docs.

[tk0miya]

Hmm, it seems the docname was wrong.
We have to replace current_source by self.env.docname like 00f4760 did.

[tk0miya]

Note: this is a bugfix. so please rebase this to 1.8 branch. then I will merge this to 1.8.5 release.
Thanks,

[lamby]

Rebased against 1.8 branch.

[mitya57]

@lamby Please also edit this pull request and change target branch from master to 1.8.

[lamby]

Ah, thanks. No wonder it was still saying "conflict"...

[mitya57]

We fix this by limiting how much of the docname variable ends up in the final constructed filename; I assume there is a good reason for including the options dictionary in the first place, otherwise we could simply omit it.

I want to note that your patch actually omits all options except the filename. Though in my opinion it is fine, as the resulting image depends only on source file name and graphviz arguments, which are present in the hash anyway.

[lamby]

the resulting image depends only on source file name

Nod. Note that it cannot include include the absolute path to the source file name, otherwise it would be unreproducible as before. ;-)

[tk0miya]

As I commented above, the value of docname must be wrong. So this is not only "not reproducible", but also a bug.
So we need to fix the value itself. self.env.docname is more appropriate for this case.

sphinx/sphinx/ext/graphviz.py

Lines 199 to 201 in 12c6149

	node['options'] = {
	'docname': path.splitext(self.state.document.current_source)[0],
	}

[lamby]

As I commented above, the value of docname must be wrong.

Don't follow that, sorry...

[tk0miya]

Merged #6111 instead.
Thank you for reporting :-)